Spatio-temporal feature extraction model based on dynamic identification of operating conditions for sintering tumbler strength prediction

Iron ore sintering is a key process for providing high-quality sintered ore for blast furnace ironmaking and the tumbler strength is an important physical indicator for measuring the quality of sintered ore. Accurately predicting tumbler strength is crucial for ensuring the efficiency of blast furnace operations and enhancing the quality of sintered ore products. This article proposes a spatio-temporal feature extraction model based on dynamic identification of operating conditions for sintering tumbler strength prediction. Firstly, in response to the complex operating conditions during the sintering process, effective learning of spatio-temporal dependencies in the data is employed to identify operating conditions in the feature space. Then, to address the imbalance in sample data across various operating conditions, the proposed framework employs a stacked spatio-temporal feature processing module to perform multi-level spatio-temporal feature extraction and fusion for each condition independently, this approach enables the model to effectively capture and learn the distinct interactions between features across different operating conditions. Additionally, a specially designed balanced loss function is integrated to optimize the model’s performance by assigning higher weights to less frequent conditions, ensuring a more equitable learning process across all operating conditions. Finally, to address potential missing data issues in the sintering process, this paper introduces a historical data pattern matching module. By matching similar historical patterns for prediction, this module helps smooth the final prediction results, thereby reducing the impact of missing data. In the end, the prediction results of the model are composed of the spatio-temporal feature processing module and the historical pattern prediction results. Compared to the baseline models, the proposed model demonstrates outstanding performance in multi-step tumbler strength prediction tasks, achieving a single-step MAE of 0.230 and RMSE of 0.258.